Some Aspects of Dimensionality and Sample Size Problems in Statistical Pattern Recognition,
Abstract
The report is concerned with obtaining relationships between the dimensionality of a measurement vector and the number of training samples available in order to maximize the performance of a pattern classifier. In statistical pattern classification, it is known that, in general, if the class-conditional densities are not completely known and only a finite number of training samples are available, then above a certain number of measurements, the performance starts deteriorating rather than improving steadily. Previous investigators have studied conditions under which this 'curse of finite sample size' can be escaped and other properties of the optimal measurement complexity. However, all these results are valid only for the specific structural assumptions and optimal estimators and generalization to other situations had to be made heuristically. In the report, several rather general results pertaining to independent measurements and arbitrary estimators are derived. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 01, 1973
- Accession Number
- AD0774121
Entities
People
- Anil K. Jain
Organizations
- Ohio State University